Sinkhorn attacks

Abstract

Adversarial attacks have been extensively investigated in the recent past. Quite interestingly, a majority of these attacks primarily work in the lp space. In this work, we propose a novel approach for generating adversarial samples using Wasserstein distance. Existing Wasserstein distance-based works generate adversarial samples using balanced optimal transport (OT). However, balanced OT requires input marginals to be of the same total probability masses these precluding its immediate application to images. Motivated by the recent unbalanced OT theory, we propose a UOT based adversarial threat model with relaxed marginal equality constraints. Our experiments on retrieval and classification tasks demonstrate significantly stronger attacks with better image quality as well as less computational overhead.